SKY-001 vs GPT-6 Astra: CAD Benchmarks, Cost and Availability

SKY-001 is Skyfall AI's "engineering world model" for CAD and building design, announced October 19, 2026. On Skyfall's own CAD test it matches GPT-6 Astra's accuracy (79.7% vs 79.1% IoU), runs about 51 times faster and costs about 150 times less per part. You can't use it yet: access is a waitlist.

For CAD work, three things matter: are the parts as good, what does each one cost, and can you use it? Going by what Skyfall has published: about as good, a fraction of the cost, and not yet.

Short answer

• Same accuracy, far cheaper, on Skyfall's test. 79.7% vs 79.1% IoU on 1,000 CAD parts, at 3.9 seconds and $0.005 a part against 199 seconds and $0.75. On 1,000 floor plans SKY-001 leads 43.6% to 40.6%.
• Nobody outside Skyfall can check it yet. The test set, the prompts and the cost math aren't public, and Skyfall's own landing page and blog list different rival models.
• SKY-001 is waitlist-only, and its output isn't an editable CAD model yet. There's no API and no price; today it gives you a visual design plan, and Skyfall lists editable CAD history as its next step. GPT-6 Astra is on the API now, and OpenAI charges $10 per million input tokens and $50 per million output tokens.
• For CAD work today, have GPT-6 Astra write the script. In our run it wrote a 132-line Blender script for a simple part in 59 seconds, about 6 cents of output at OpenAI's price.

What is SKY-001, and how is it different from GPT-6 Astra?

SKY-001 plans a build the way an engineer does, while GPT-6 Astra writes code that you run.

Skyfall calls it a "world model" that "doesn't do next-token prediction like LLMs". Its Hierarchical Composer splits an object into a part tree; its Construction World Model then draws each part being built, sketch by sketch and 3D step by step.

For CAD you give it an image, text or a rough sketch; for BIM, a storey outline and a short brief.

The output is a "visual design plan": the part tree plus the sketches and 3D states for each part, viewable as a 3D assembly. The press release says it can be viewed in CAD software such as Autodesk Fusion. Skyfall lists "executable operations with an editable feature history" as its next step, so today you get a plan to look at.

OpenAI calls GPT-6 Astra its most capable model. It was announced September 3, 2026, and its model ID is gpt-6-astra. It takes text and images and returns text; for CAD you ask it for a script, such as Blender Python (bpy), and run that script yourself.

Skyfall AI says it was founded by the team behind Maluuba, the deep-learning lab Microsoft acquired. Sam Pasupalak (co-founder and CEO), Kaheer Suleman and Sumit Pasupalak lead it from San Francisco, with a research hub in Toronto. Its backers are Fidelity, Touring Capital, M13 and iNovia Capital; no amount is stated.

SKY-001 learned from more than 200,000 parts and over 600,000 construction examples for CAD, plus about 35,000 floor plans for BIM.

The practical difference today: GPT-6 Astra gives you a script you can run, while SKY-001 gives you a build plan to look at, once you get access.

What do Skyfall's CAD and BIM numbers show?

On Skyfall's own tests, SKY-001 ties GPT-6 Astra on CAD accuracy, edges it on floor plans, and is far faster and cheaper on both.

The CAD test samples 1,000 parts from complex objects such as a Jaguar E-Type, a FANUC robot arm and a jet engine. Each part takes about 100 drawing steps. The score is IoU: how much the built shape overlaps the reference, where 100% is a perfect match.

The BIM test uses 1,000 floor plans, about 350 whole floors and 650 apartments, at about 70 drawing steps each on average. It is scored by IoU too.

These are Skyfall's numbers, from its SKY-001 page and its October 19, 2026 blog post.

For 1,000 CAD parts, that is $5 against $750, and about 65 minutes against 55 hours if you run them one after another. On floor plans the gap is bigger still: about 119 times faster and 186 times cheaper.

Among the language models in Skyfall's tables, GPT-6 Astra scores highest on CAD. On floor plans, Gemini 3.8 Flash beat it, 42.3% to 40.6%, at about half the cost per plan ($0.218) but more than twice the time (346.2 s). Claude Opus 5.5 scored 69.4% on CAD at $0.31 a part (landing page only).

Skyfall also tested whole assemblies. From one image, GPT-6 Astra, Claude Opus 5.5 and Claude Fable 5.1 each wrote a single Blender script for the full assembly, and Skyfall says GPT-6 Astra was far ahead of the two Claude models.

Two AI judges, GPT-6.1 Sol and Claude Opus 5.5, then scored SKY-001 against GPT-6 Astra without knowing which made which. SKY-001 scored higher on all five criteria, for example 8.1 vs 6.7 out of 10 on matching the input image, and GPT-6 Astra returned a model for only 82% of the objects.

So if Skyfall's numbers hold, the real gap is speed and cost; on CAD accuracy, the two are close.

What can't you verify yet?

Almost none of it, because Skyfall hasn't published the test set, the prompts or the cost math.

1. No public test set, paper or code; the blog only offers a newsletter for future papers and benchmarks.
2. Skyfall doesn't say how it measured time and cost: which prices, which reasoning setting, which hardware.
3. Skyfall's two pages disagree on the rivals. On CAD, the blog shows Gemini 4 at 76.9% and no Claude; the landing page shows Gemini 3.8 Flash at 54.7% and Claude Opus 5.5 at 69.4%. The SKY-001 and GPT-6 Astra rows are the same on both pages.
4. The assembly scores come from AI judges, not people or measurements.
5. The output isn't an editable CAD model yet. Skyfall's own caption on one build replay says the saved sketch and 3D geometry "can disagree", and editable feature history is listed as future work.

OpenAI has its own CAD number, from a different test. On BenchCAD, where a model rebuilds a 3D object from rendered views by writing CAD code, OpenAI reports 95.9% with tools, against 83.3% for GPT-5.6 Sol and 84.3% for Claude Fable 5.1. The test and the scoring differ, so 95.9% can't be set against Skyfall's 79.1%.

What would settle it is a public test set anyone can rerun. Until then, treat "150 times cheaper" as Skyfall's claim, not an independent result.

Can you use SKY-001 or GPT-6 Astra today?

GPT-6 Astra yes, SKY-001 no.

Those prices are per million tokens.

Skyfall's FAQ says "Join the waitlist to be the first to hear when it goes public". The press release calls SKY-001 an initial research preview, with full commercial deployment targeted for early 2027.

OpenAI charges $10 per million input tokens ($1 for cached input) and $50 per million output tokens. A request with more than 272K input tokens is billed at 2x input and 1.5x output for the whole request; Batch and Flex cost half.

The context window is 1,050,000 tokens, a reply can run to 128,000 tokens, and reasoning effort can be low, medium, high, xhigh or max.

OmniaKey serves gpt-6-astra through an OpenAI-compatible API at https://api.omniakey.com/v1. The same key also reaches other models from Skyfall's tables, such as Claude Opus 5.5 and Gemini 3.8 Flash. Rates are on the GPT-6 Astra page.

Which should you use for CAD work right now?

GPT-6 Astra, writing scripts you run in Blender or your CAD tool. Join the SKY-001 waitlist if you build complex assemblies at volume.

We sent GPT-6 Astra this request through the OmniaKey API on October 22, 2026, at medium reasoning effort:

Prompt:

Write a Blender 4.2 Python script (bpy and bmesh only, no add-ons) that builds this machined part as one mesh object named "BearingBlock". Units are millimetres: set the scene so 1 Blender unit = 1 mm.

1. Base plate: 120 × 40 × 12 mm, centred on the origin in X and Y, bottom face at Z = 0.
2. Two through holes, diameter 11 mm, in the base plate at X = -45 mm and X = +45 mm (Y = 0).
3. Boss: a vertical cylinder, diameter 56 mm, standing on the plate from Z = 12 to Z = 56, centred at X = 0, Y = 0.
4. Bore: a through hole, diameter 35 mm, along the boss axis, through the boss and the plate.
5. Use Boolean modifiers (EXACT solver) and apply them, so the result is one closed mesh.
6. At the end, print the object's dimensions and vertex count.

Reply with the complete script in one Python code block and nothing else.

• It returned a 132-line Blender script in 59 seconds, with 1,117 output tokens, about 6 cents at OpenAI's $50 per million.
• The script used every dimension we asked for and Blender's exact Boolean solver, and it checks that the mesh is closed.
• A second run of the same request through the Python code below took 100 seconds and 1,273 output tokens.
• We checked the script by reading it, not by running it in Blender.

Skyfall's CAD cases take about 100 steps each, so its 199 seconds and $0.75 for GPT-6 Astra describe harder jobs than this one-part example.

import os
from openai import OpenAI

client = OpenAI(
    base_url="https://api.omniakey.com/v1",
    api_key=os.environ["OMNIAKEY_API_KEY"],
)

prompt = open("part.md").read()  # the part description above

response = client.chat.completions.create(
    model="gpt-6-astra",
    reasoning_effort="medium",
    messages=[{"role": "user", "content": prompt}],
)

print(response.choices[0].message.content)  # the Blender script
print(response.usage)  # tokens, to estimate the cost

Install the OpenAI package with pip install openai, save the prompt as part.md and set OMNIAKEY_API_KEY. Run the code, paste the script from the reply into Blender's Scripting tab, and click Run Script. Our GPT-6 Astra setup guide covers Node.js, the Responses API and Codex.

At a bigger scale, OpenAI's developer blog shows GPT-6 Astra building an editable house scene through the Blender Python API, then refining it from renders.

Pick by what you need:

• You need parts or scenes now → GPT-6 Astra writing Blender or CAD scripts.
• You make many floor plans and cost matters more than speed → Gemini 3.8 Flash scored higher on Skyfall's floor-plan test at about half the cost, though it is slower.
• You build complex assemblies at volume and can wait → join the SKY-001 waitlist. Skyfall targets early 2027; check that it exports editable CAD before you plan around it.

Common questions

Is SKY-001 open source?

No. Skyfall has published no code, weights, paper or licence; access is a waitlist.

Is there a SKY-001 API or price?

Not yet. Skyfall's pages list no public API and no price, and commercial deployment is targeted for early 2027.

Who is behind Skyfall AI?

Skyfall AI says it was founded by the team behind Maluuba, the deep-learning lab Microsoft acquired. Sam Pasupalak (CEO), Kaheer Suleman and Sumit Pasupalak lead it from San Francisco, with a research hub in Toronto.

Is GPT-6 Astra good at CAD?

It scores highest among the language models in Skyfall's CAD table, at 79.1%. OpenAI reports 95.9% on BenchCAD with tools, a different test with different scoring.

Can GPT-6 Astra make Blender models?

Yes, by writing Blender Python scripts that you run. OpenAI's house demo and our own run both work that way.

Sources

• Skyfall: SKY-001 page and blog post (October 19, 2026)
• Skyfall press release
• OpenAI: GPT-6 Astra announcement (September 3, 2026)
• OpenAI: GPT-6 Astra model page and pricing
• OpenAI developer blog: architectural visualization with Astra

All checked on October 22, 2026.